The Missing Verb
An AI executive stood in the Vatican and asked for critics his industry's incentives could not bend. Australia's answer, so far, is a rule about water.
On 15 July, Anthony Albanese stood up in the Great Hall at the University of Sydney and gave a speech called AI in Australia's interests. He announced an Office of AI inside the Department of the Prime Minister and Cabinet, effective that day, and a set of Australian Standards for AI to pull together a response that had until then been handled sector by sector. The Government says it will be the first legislated national AI framework anywhere. National Cabinet considers it in August, with legislation to follow early next year. August is this month, which is why I am writing now rather than in October, when the interesting part will be over.
What sat behind the speech came out separately, under freedom of information law. In April, Dario Amodei, chief executive of the AI company Anthropic, met the Treasurer, Jim Chalmers. The briefing notes show Amodei had asked for the meeting to discuss barriers to AI training in Australia, copyright reform in particular, and that officials expected him to say investment in model development and the data centres behind it was contingent on clarity of copyright settings.
Nobody misbehaved here. Lobbying a government about the legal conditions of a very large investment is lawful, ordinary, and in this case written down and then released. Anthropic has been franker than most of its competitors about what it wants and why, which is what makes the episode useful rather than grubby. According to Capital Brief, the Australian Financial Review reported that the company had gone to five Australian data centre operators asking for at least 1.4 gigawatts, roughly the whole existing Australian industry, in a deal worth up to A$21.6 billion. The Government refused a text and data mining exception. Albanese said Australian work cannot be used to train AI without the artist's control.
Six weeks before all this, a co-founder of that same company stood in a room in Rome, in front of the Pope and his cardinals, and said his own industry cannot be trusted to police itself.
Christopher Olah leads the research team at Anthropic that studies what actually happens inside the company's models. He had been invited to speak at the presentation of Magnifica humanitas, the first encyclical of Pope Leo XIV, subtitled "On Safeguarding the Human Person in the Time of Artificial Intelligence". The Pope presented it himself, which popes almost never do. Olah began by warning the room that what followed might "sound strange coming from the co-founder of an AI company". Every frontier lab, his own included, sits inside incentives that pull against doing the right thing. Commercial survival. Staying at the research frontier. Geopolitics. Then the older pressures, pride and ambition. However decent the people involved (and he believes many of them are), the pull is still there. So the technology only goes well, he argued, if somebody outside those incentives is paying attention and willing to say the hard thing. He asked for moral voices that the incentives could not bend.
Read that again, because it has nothing to do with faith. He is making a claim about how institutions are built: a system cannot audit itself, and the correction has to come from a party whose interests differ. Any company secretary would recognise the argument and that is the reason we invented boards.
He said it in Rome, to clergy. Six weeks later Australia gave him the worked example, on schedule, inside our copyright policy, with our grid attached to it.
So his question is now our question. In this country, who has interests different enough to say the hard thing?
Look at where Australian policy reaches for binding language.
The announced data centre standards are precise. Underwrite your own new power supply. Pay your full share of connection costs so household bills are not hit. Cut your draw when the grid needs it. Use as little water as you can. That is a shall, and a good one, though for now it is an announcement rather than a statute.
The public service has its own, and that one is already live. The Digital Transformation Agency's revised Policy for the Responsible Use of AI in Government took effect on 15 December 2025 and binds every non-corporate Commonwealth entity outside defence and the intelligence community. Keep a register of AI use cases with a named owner for each. Run an impact assessment before you deploy. Have a process for AI incidents. Give staff and the public a route to report safety concerns. Train everyone on the payroll. The first of the new mandatory requirements bit on 15 June this year, the rest arrive in December, and agencies had to appoint a Chief AI Officer by the end of June.
Now the private employer. A company that ranks job applicants by model, or scores productivity through a monitoring tool, or lets software decide who gets the shift, gets the National AI Centre's Guidance for AI Adoption and its six essential practices. Guidance. The National AI Plan of December 2025 confirmed Australia would lean on existing law and sector regulators rather than impose mandatory guardrails across the economy, after the Productivity Commission argued that guardrails would chill innovation.
The standard reassurance from employment lawyers is that the old law already covers it. If an algorithm fires you, the employer still needs a valid reason and a fair process, and discrimination law still applies to a model that screens on age or disability. All true. Now go looking for the Australian cases. I have, and I cannot find them. What the Fair Work Commission is dealing with is the reverse problem, workers using generative AI to draft their claims, which its president Justice Adam Hatcher told the Victorian Bar Association in February has helped push the Commission's workload up around 70% in three years. On algorithmic management of employees, the literature is all conditional. Even if an algorithm made the decision, the employer would remain liable. Would.
The absence is the point. Nothing requires an Australian private employer to record which systems it runs, assess them before deployment, or tell anyone they exist. So the harm does not surface as a case. It surfaces as a resignation, or a WorkCover claim, or a roster nobody can explain, or as nothing at all. The first real Australian decision on this will not arrive because a company disclosed something. It will arrive because someone found out.
Put the three tiers next to each other and the shape is hard to miss. Our binding language covers megawatts, litres of water, grid connection costs and the conduct of Commonwealth public servants. Our advisory language covers what a private company may do to the people who work for it.
I am not taking a shot at the Government, which has moved faster than most comparable countries and held its ground on copyright while a very large cheque sat on the table. The sequencing makes sense to me. A megawatt is easier to define than a fair dismissal.
But look at what the July announcement actually named. Location controls for data centres. Power. Water. Ownership and consent for artists, writers, musicians and journalists. Nothing about what a company may do to its own staff with a model. The rest of the framework's scope goes to industry consultation, so nobody can yet tell a board whether the standards will ever reach that far. Which means the honest way to describe this gap is not that it closes in 2027. It is that no date has been put on it at all, and boards have been quietly assuming otherwise.
Which brings me back to the encyclical, worth reading even if the Church means nothing to you. Magnifica humanitas argues that no technology arrives neutral, because it carries the marks of whoever built it, paid for it, regulated it and used it. It says AI promises to lift productivity by absorbing dull tasks and then, in practice, makes workers speed up to match the machine instead of the machine being built around the worker.
Any of that could go into a board charter tomorrow. None of it is written as an obligation. The encyclical says "should, urges, calls for, hopes", and it has no other option available to it. A pope writes to consciences. He has no register, no penalty, no compliance team, no power to stop a system going live on Monday. Olah asked for critics the incentives cannot bend. The Church accepted and did the analysis. What it does not have is the verb.
One body in Australian corporate life holds the verb and leaves it in the drawer.
A board can resolve that the company shall keep a register of its AI systems with a named owner against each one. That no system touching somebody's job, credit, safety or access to a service shall go live without a documented impact assessment. That there shall be a way for an employee or a customer to challenge an automated outcome and reach a human being who can overturn it. That AI incidents shall come to the board the way safety incidents do. That management shall report against all of it quarterly, in writing, with the seriousness we give a lost time injury frequency rate.
None of this needs inventing. An Australian government agency wrote the instrument, the Commonwealth has been phasing it in since June, and it is published free.
Here is the objection I would expect from the other side of the table, and it deserves an answer. Adopt a standard you are not legally required to meet, and you have handed a future plaintiff a yardstick. Every impact assessment you write is a document that can be produced. Every incident you log is an admission that the incident happened. Directors who have been through a class action know exactly how that feels.
My answer is that the yardstick exists whether you write it down or not. The Fair Work Act, the discrimination statutes and the WHS duties already apply to a decision your software made, and the day someone tests that, the question in the room will be what you knew and when. A board that ran no register will have nothing to say. Documented process has been a defence in safety law for decades rather than a liability, and the companies that fought hardest against writing things down are the ones that ended up explaining themselves without records. The narrower point is that the DTA policy is the closest thing Australia has to a settled standard here, so a board adopting it is borrowing the government's homework rather than guessing.
Why the board and not the executive? For the reason Olah gave in Rome. Management sits inside the incentives, carrying the productivity target, the cost-out programme, the deployment date, and often a bonus tied to all of them. That is their job, and I would not want it done by someone indifferent to any of it. The board is the one organ of a company built to sit next to those incentives instead of inside them, with duties running to the company rather than to this year's numbers. Directors are, in the plain and unglamorous sense Olah meant, the critics the incentives are not supposed to bend.
Whether they are is a different question, and I have sat in enough rooms to be unsure.
In Western Australia the whole thing lands harder. National Cabinet is weighing a framework whose sharpest obligations are about power and water, in a state whose South West Interconnected System has no neighbour to borrow from. Perth boards will get the megawatt shall whether they wanted it or not, and nothing binding at all about the software allocating shifts or ranking who goes on the redundancy list. Most of the national commentary will be written in Sydney and Melbourne, where the grid is an abstraction everyone shares and a mine site is a line in an annual report.
Olah closed by saying the day was a beginning, and that what was needed was more of the world looking closely and pushing things in a better direction. He was talking to the Church, which will do its part slowly, in its own register, over decades. There were no Australian directors in the room, and on this particular question nobody else is coming to write the missing verb.